DeepSeek

DeepSeek-V4.1-Flash

DeepSeek

DeepSeek-V4.1-Flash is a natively multimodal MoE model trained from scratch on 45 trillion tokens. It uses a causal encoder-decoder layout, Compressed Sparse Attention 2, Engram conditional memory, and DSpark speculative decoding to reduce inference and KV-cache cost. Its backbone activates different parameter counts during prefill and decoding, while a separate 196B conditional-memory table is sparsely accessed. The model accepts images and text and generates text, supports million-token contexts and adjustable reasoning effort, and powers the official deepseek-flash API. MIT weights and model-specific inference code support self-hosting.

DeepSeek-V4-Flash-Vision-Exp

DeepSeek

DeepSeek-V4-Flash-Vision-Exp is the experimental vision-language specialization of the production V4 Flash model. It adds image understanding for agent tasks while retaining the text capabilities of V4 Flash, enabling workflows that inspect interfaces, diagrams, and other visual context. It was publicly accessible through a dedicated DeepSeek API model name before downloadable MIT weights were released. Following the September 10 V4.1 Flash launch, the experimental endpoint was retired and its alias redirected; the original checkpoint remains available for self-hosted research.

DeepSeek-V4-Flash

DeepSeek

DeepSeek-V4-Flash is the smaller, efficiency-focused tier of the V4 family, first publicly released as a preview on April 24, 2026. Its hybrid Compressed Sparse Attention and Heavily Compressed Attention architecture supports million-token contexts alongside manifold-constrained hyper-connections. It supports non-thinking and thinking modes for code, knowledge, and agent tasks. The 2026-07-31 production update (DeepSeek-V4-Flash-0731) adds substantially stronger agent behavior and DSpark speculative decoding, and is included as a named checkpoint variant. The original Flash API has retired and its alias now routes to V4.1 Flash; original and updated V4 weights remain downloadable.

DeepSeek-V4-Pro

DeepSeek

DeepSeek-V4-Pro is the flagship tier of the V4 family, first publicly released as a preview on April 24, 2026. Its hybrid Compressed Sparse Attention and Heavily Compressed Attention architecture supports million-token contexts alongside manifold-constrained hyper-connections. It supports non-thinking and thinking modes for code, knowledge, and agent tasks. The 2026-08-13 production update (DeepSeek-V4-Pro-0813) adds substantially stronger agent behavior and DSpark speculative decoding, and is included as a named checkpoint variant. The Pro API remains available after the V4.1 Flash launch.

DeepSeek-V3.2-Speciale

DeepSeek

DeepSeek-V3.2-Speciale is the high-compute reasoning specialization released alongside V3.2. A more extensive reinforcement-learning phase targets difficult mathematical and scientific problems, trading generation efficiency for reasoning depth. The downloadable checkpoint remains available under MIT terms. Its official temporary API endpoint expired on December 15, 2025, and did not support tool calls. Speciale should therefore be treated as a distinct reasoning model with a different capability and access profile from the ordinary V3.2 agent model.

DeepSeek-V3.2

DeepSeek

DeepSeek-V3.2 combines DeepSeek Sparse Attention with scalable reinforcement learning and large-scale synthetic agent training. It supports both ordinary responses and explicit reasoning, including tool use within thinking mode. The release emphasizes mathematical reasoning, software engineering, and general agent workflows while reducing long-context inference cost. Public MIT weights permit self-hosting. DeepSeek originally exposed it through the deepseek-chat and deepseek-reasoner aliases, which subsequently migrated to V4; the model card and weights identify the specific release. Speciale, the separate high-compute reasoning specialization, has its own page.

DeepSeekMath-V2

DeepSeek

DeepSeekMath-V2 is a large open-weight model for self-verifiable mathematical reasoning. It moves beyond final-answer rewards by training the model to assess proof quality and refine its solutions, combining a proof generator with learned verification and iterative improvement. It is intended for difficult competition-style mathematical proofs, including olympiad problems. The checkpoint is released under Apache-2.0 with research and inference materials. Its proof-generation focus is distinct from both general DeepSeek reasoning models and Lean-based formal provers.

DeepSeek-OCR

DeepSeek

DeepSeek-OCR is a vision-language model for converting visually encoded documents into text. The initial release investigates optical compression of long textual contexts, combining a compact visual encoder with a language decoder for OCR, structured extraction, and document-to-Markdown conversion. Public weights, code, and local inference examples support self-hosted document processing. Its image-to-text language decoder places it in scope as a specialized language-capable multimodal model, rather than an embedding or image generator. The official README also demonstrates general image description using a natural-language prompt, establishing language-capable visual understanding beyond pure transcription.

DeepSeek-V3.2-Exp

DeepSeek

DeepSeek-V3.2-Exp is an independently released experimental update introducing DeepSeek Sparse Attention (DSA). It tests a more efficient attention mechanism for long-context inference while retaining the thinking and non-thinking modes of the V3.1 family. Base and post-trained weights are available under MIT terms. This preview was deployed on DeepSeek services before the final V3.2 release; the current API aliases no longer select it. Its primary significance is the sparse-attention architecture and deployment-cost experiment, rather than a new parameter tier.

DeepSeek-V3.1

DeepSeek

DeepSeek-V3.1 unifies thinking and non-thinking behavior in one open-weight model. Continued pretraining and agent-focused post-training improve software engineering, search, and tool invocation while making reasoning more efficient than R1-0528. The weights include base and post-trained checkpoints. The September 22 Terminus update refines language consistency and code/search agents without changing the underlying model size; it is represented here as a variant. Historical deepseek-chat and deepseek-reasoner aliases have since migrated to newer models.

DeepSeek-R1-0528

DeepSeek

DeepSeek-R1-0528 is a substantial post-training update to the R1 reasoning model, improving mathematics, programming, and general reasoning through more inference-time computation. DeepSeek reports lower hallucination rates, better front-end generation, and support for JSON output and function calling. The full MoE checkpoint is released alongside a Qwen3-8B distilled variant. Its historical deepseek-reasoner API alias has since moved to newer models; downloadable weights preserve this release. Longer reasoning traces can consume more tokens than the original R1.

DeepSeek-Prover-V2

DeepSeek

DeepSeek-Prover-V2 is a family of language models specialized for formal theorem proving in Lean 4. It uses subgoal decomposition and reinforcement learning to convert difficult proof tasks into verifiable intermediate steps. The 671B checkpoint is trained from DeepSeek-V3-Base, while the 7B checkpoint builds on Prover-V1.5-Base and supports a 32K context. The published model repository provides weights, inference instructions, and proof benchmarks. Formal validity depends on checking generated code in the Lean environment and matching its theorem statement.

Janus-Pro

DeepSeek

Janus-Pro unifies language-based multimodal understanding and image generation in a single autoregressive Transformer. It separates the visual encoding pathways for image understanding and generation to reduce conflicts between the two objectives. The public 1B and 7B models answer image-grounded questions and generate images from text, so they qualify as language-capable multimodal models. This page represents both routine size tiers. The model card distinguishes MIT-licensed code from the custom DeepSeek Model License governing the weights.

DeepSeek-R1

DeepSeek

DeepSeek-R1 is a first-generation reasoning model trained via large-scale reinforcement learning. Built on DeepSeek-V3-Base, it incorporates cold-start data before RL to address challenges like endless repetition and poor readability found in DeepSeek-R1-Zero. Achieves performance comparable to OpenAI-o1 across math, code, and reasoning tasks through advanced chain-of-thought reasoning capabilities.

DeepSeek-V3

DeepSeek

A powerful Mixture-of-Experts (MoE) language model with 671B total parameters and 37B activated for each token. Features Multi-head Latent Attention (MLA) and DeepSeekMoE architectures with innovative auxiliary-loss-free load balancing and multi-token prediction training. Pre-trained on 14.8T high-quality tokens with only 2.788M H800 GPU hours.

DeepSeek-VL2

DeepSeek

DeepSeek-VL2 is a vision-language MoE family for image-grounded conversation, OCR, visual grounding, and document, chart, and table understanding. Dynamic image tiling supports higher-resolution inputs, while the language backbone uses Multi-head Latent Attention. The Tiny, Small, and flagship checkpoints activate roughly 1.0B, 2.8B, and 4.5B language parameters respectively. Downloadable checkpoints and demonstration code enable local experimentation. Model weights use the DeepSeek Model License; this page groups the routine size variants rather than creating duplicate family pages.

DeepSeek-V2.5

DeepSeek

DeepSeek-V2.5 combines the general conversation abilities of V2-Chat with the programming specialization of Coder-V2-Instruct. It strengthens instruction following, writing, code generation, and human-preference alignment in a single post-trained model. The V2.5-1210 checkpoint further updates mathematics, coding, and writing and is grouped as a variant. The original API accepted both chat and coder aliases, which now serve later generations; the named weights remain downloadable under the custom DeepSeek Model License.

DeepSeek-Coder-V2

DeepSeek

DeepSeek-Coder-V2 is an open-weight coding MoE family built by continuing pretraining from DeepSeek-V2 on additional code and mathematics data. It supports 338 programming languages, repository-scale context, and fill-in-the-middle completion. The family includes 236B and 16B Lite checkpoints, each with base and instruct versions. The 128K context specification applies to the published family. Downloadable weights use the DeepSeek Model License; historical deepseek-coder API access has migrated to later models.

DeepSeek-V2

DeepSeek

DeepSeek-V2 introduced DeepSeekMoE and Multi-head Latent Attention to reduce the cost of storing attention states and running a large language model. Its flagship has 236B parameters with 21B activated per token and supports 128K contexts. The release includes base/chat models and a smaller V2-Lite tier. It is a historical foundation for the V2.5 and later DeepSeek families; downloadable weights remain available under the DeepSeek Model License. Its Chinese and English evaluations cover knowledge, mathematics, coding, and instruction following.

DeepSeekMath

DeepSeek

DeepSeekMath is a 7B mathematics-focused language-model family built through continued pretraining of DeepSeek-Coder-Base-v1.5 on mathematical web content, followed by instruction tuning and reinforcement learning. Its paper introduces Group Relative Policy Optimization (GRPO), an approach subsequently used in reasoning-model research. Base, Instruct, and RL checkpoints are available. The model targets mathematical problem solving and structured reasoning rather than an all-purpose multimodal assistant. Weights are distributed under the custom DeepSeek Model License, with commercial use supported under its terms.

DeepSeek Coder

DeepSeek

DeepSeek Coder is a family of code-specialized decoder-only language models trained on code and natural language. Project-level training and fill-in-the-middle objectives support code completion, while instruction-tuned releases target code generation and explanation. The original release spans 1.3B to 33B parameters, with base and instruct checkpoints. This page covers the original coding family, including the later 7B v1.5 checkpoint; the substantially different Coder-V2 architecture has its own page. Weights permit commercial use subject to the DeepSeek Model License.

DeepSeek LLM

DeepSeek

DeepSeek LLM is an advanced language model comprising 67 billion parameters, trained from scratch on a vast dataset of 2 trillion tokens in both English and Chinese. The model demonstrates superior general capabilities, outperforming Llama2 70B Base in reasoning, coding, math, and Chinese comprehension. Available in both Base and Chat variants with 7B and 67B parameter sizes.